wiki3-ai

Python wrapper for Chrome's built-in AI Prompt API, enabling direct access to browser-provided language models from Jupyter notebooks.

Features

• 🤖 Direct access to Chrome's built-in AI via the Prompt API
• 📓 Seamless Jupyter integration using AnyWidget
• 🔄 Bidirectional communication via traitlets
• 🎯 Type-safe Python API matching the Web IDL specification
• 🌊 Support for streaming responses
• 🛠️ Tool use capabilities
• 🖼️ Multimodal input support (text, images, audio)
• 📊 Session management and cloning
• 🔒 Privacy-focused on-device processing

Installation

pip install wiki3-ai

Requirements

• Python 3.10+
• Chrome browser with Prompt API enabled
• Running in a Jupyter environment (JupyterLab, Jupyter Notebook, etc.)

Quick Start

from wiki3_ai import LanguageModel

# Create a language model session
session = await LanguageModel.create()

# Simple prompt
result = await session.prompt("Write me a poem about Python.")
print(result)

# Streaming response
async for chunk in session.prompt_streaming("Write me a long story."):
    print(chunk, end="", flush=True)

Usage Examples

System Prompts

from wiki3_ai import LanguageModel, LanguageModelMessage, LanguageModelMessageRole

session = await LanguageModel.create({
    "initialPrompts": [
        {
            "role": "system",
            "content": "You are a helpful Python programming assistant."
        }
    ]
})

response = await session.prompt("How do I read a file in Python?")
print(response)

Checking Availability

from wiki3_ai import LanguageModel, Availability

# Check if the API is available
availability = await LanguageModel.availability()
print(f"Model availability: {availability}")

if availability == Availability.AVAILABLE:
    session = await LanguageModel.create()
    # Use the session...

Configuring Temperature and Top-K

# Get default parameters
params = await LanguageModel.params()
print(f"Default temperature: {params.default_temperature}")
print(f"Max temperature: {params.max_temperature}")

# Create session with custom parameters
session = await LanguageModel.create({
    "temperature": 0.8,
    "topK": 40
})

Session Management

# Create a session
session = await LanguageModel.create()

# Use the session
result1 = await session.prompt("Tell me about AI.")

# Clone for different conversation branches
session2 = await session.clone()
result2 = await session2.prompt("Now tell me about machine learning.")

# Destroy when done
await session.destroy()

Measuring Token Usage

session = await LanguageModel.create()

# Check current usage
print(f"Current usage: {session.input_usage}/{session.input_quota}")

# Measure potential usage before prompting
usage = await session.measure_input_usage("This is my prompt")
print(f"This prompt would use: {usage} tokens")

# Prompt if there's enough quota
if session.input_usage + usage < session.input_quota:
    result = await session.prompt("This is my prompt")

Structured Output with JSON Schema

schema = {
    "type": "object",
    "required": ["rating"],
    "properties": {
        "rating": {
            "type": "number",
            "minimum": 0,
            "maximum": 5
        }
    }
}

result = await session.prompt(
    "Rate this: The food was excellent!",
    {"responseConstraint": schema}
)

import json
data = json.parse(result)
print(f"Rating: {data['rating']}")

API Reference

LanguageModel

Main class for interacting with the language model.

Class Methods

• create(options=None) - Create a new language model session
• availability(options=None) - Check model availability
• params() - Get model parameters

Instance Methods

• prompt(input, options=None) - Send a prompt and get response
• prompt_streaming(input, options=None) - Send a prompt and stream response
• append(input, options=None) - Append messages without getting response
• measure_input_usage(input, options=None) - Measure token usage
• clone(options=None) - Clone the session
• destroy() - Destroy the session

Properties

• input_usage - Current token usage
• input_quota - Maximum token quota
• top_k - Top-K sampling parameter
• temperature - Temperature sampling parameter

Data Models

• LanguageModelMessage - A message in the conversation
• LanguageModelMessageContent - Content with type and value
• LanguageModelMessageRole - Message role (system/user/assistant)
• LanguageModelMessageType - Content type (text/image/audio)
• LanguageModelCreateOptions - Options for creating sessions
• LanguageModelPromptOptions - Options for prompting
• LanguageModelParams - Model parameters
• Availability - Availability status enum

Architecture

This package uses:

• AnyWidget for Jupyter integration
• Traitlets for bidirectional Python ↔ JavaScript communication
• Chrome Prompt API for accessing built-in language models

The communication flow:

1. Python code calls methods on LanguageModel
2. Requests are serialized via traitlets to JavaScript
3. JavaScript calls Chrome's native Prompt API
4. Results are sent back via traitlets to Python
5. Python code receives async results

Specifications

This implementation follows:

• Chrome Prompt API Specification
• WebIDL Interface Definitions
• Chrome Developer Documentation

License

Apache License 2.0 - See LICENSE file for details

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

Links

• GitHub Repository
• Chrome Prompt API Documentation
• Web Machine Learning Community Group